Paper Explained: Video Generation Models: A Survey of Post-Training and Alignment — Four Toolboxes for Making Video Models Do What You Ask

A from-scratch reading of the survey that organizes how pretrained video generators are steered toward instruction-following, temporal coherence, physical plausibility, and safety without retraining—via supervised fine-tuning, self-training and distillation, preference and reward optimization, and inference-time methods.

Video Generation Models: A Survey of Post-Training and Alignment

What this paper organizes

This article reads "Video Generation Models: A Survey of Post-Training and Alignment" (Chaoyu Li et al., a collaboration led from Arizona State University; arXiv:2610.00812, posted September 30, 2026; published in TMLR 2026). It does not propose a new model. It is a survey that maps existing work into a single structure.

Here is the abstract in plain terms. Video generation has moved quickly toward high-resolution, long-duration output, yet even models with strong priors from large-scale pretraining often fail to follow human intent reliably, stay temporally coherent, or respect physical and safety constraints. Video brings problems that images and text do not: errors accumulate over time, motion and appearance are entangled, objectives trade off against each other, and supervision for temporal properties is scarce. The survey treats post-training as a unifying framework, separates implicit from explicit alignment according to how the alignment signal is enforced, and sorts existing methods into four families: (1) supervised fine-tuning, (2) self-training and distillation, (3) preference- and reward-based methods, and (4) inference-time methods. It also reviews datasets, benchmarks, and evaluation practice, and discusses open problems.

Analogy: bringing a trained cook in line with the house style

Pretraining is like a new cook who learned by watching an enormous pile of cooking videos online. They can make almost anything, but it tends to taste "average." The paper says much the same thing: pretrained models optimize data likelihood rather than human utility, so they tend to reproduce the "average" web video (§2.3).

Post-training is the apprenticeship at a specific restaurant. You can have the cook follow recipes and correct them (supervised fine-tuning), have them reproduce the head chef's results in fewer steps (distillation), let diners compare two plates and lean toward the favorite (preference and reward), or taste right before serving and add salt (inference-time methods). None of these retrains the cook from scratch.

What post-training actually adds

The paper frames video generation as learning a conditional distribution (§2.1):

In words, this says: give the model a condition c (such as a text prompt, a reference image, or an edit instruction), and it draws a video v from its learned distribution.

In words: given a condition $\mathbf{c}$ (a text prompt, reference image, or edit instruction), draw a video $\mathbf{v}$ from the model. Post-training doesn't rebuild this distribution. It tilts it toward the outputs we want. Many supervised fine-tuning methods share this form (§3.1):

In words, the first term pulls the model toward the target video, while the second acts as a brake that keeps it from destroying its pretrained abilities and temporal consistency; λ sets how strong that brake is.

The first term says "move toward the target video." The second is a brake that says "don't destroy pretrained priors or temporal consistency," and $\lambda$ sets how hard that brake is. In practice the workhorse is LoRA, which updates only a small part of the model. The survey cites PanoLora, for example, which treats panoramic video as a specialization that low-rank updates can capture (§3.2).